Justice Notes: Sentencing Guidelines and AI
A White-Collar Journal forum for criminal justice, lived experience, and the personal search for redemption
I’m publishing today an excerpt from an essay by Jonathan Wroblewski, author of Sentencing Matters, exploring a question that’s no longer hypothetical:
Could AI do the work of sentencing commissioners—better than humans?
Wroblewski has been running a series of experiments testing how large language models handle real sentencing-policy problems. In this installment, he asks two leading AI systems to resolve actual circuit conflicts identified by the U.S. Sentencing Commission—and to explain their reasoning like commissioners would.
The results are worth your time.
From Sentencing Matters:
Lots of us these days are wondering whether AI is coming for our jobs. There is already ample evidence that AI can do many jobs better, faster, cheaper than human beings, including many jobs that involve processing words, ideas, and code. So, it seems worth exploring here whether AI could be a better sentencing commissioner than most people?
This is the fourth in a series of experiments we’ve run about artificial intelligence and its place in sentencing and corrections law, policy, and practice. We will have more experiments over the coming months. Among them will be to ask various large language models to undertake the responsibilities of a sentencing commissioner and see how they do.
This experiment asks two AI models to each resolve two circuit conflicts that the U.S. Sentencing Commission published for public comment back in January and for which it has received written and oral comment this year.
The responsibilities of a sentencing commission include not only deciding among proposed guideline amendment options, but also explaining its decision. For some time, I’ve felt that the commissioners of the U.S. Sentencing Commission, in general, do not adequately explain their votes — and the reasoning behind them — on proposed guideline amendments.
It seems that it would be best practice for the Commission, like the Supreme Court, federal appellate courts, and many federal district courts too, to issue at least one detailed written opinion explaining each of its decisions…
The Experiment
Wroblewski then gave both AI systems a realistic assignment:
act as sentencing commissioners and resolve two live circuit conflicts involving the definition of “controlled substance offense.”
Each model had to:
Choose between competing legal options
Apply statutory and constitutional principles
Write a 500-word judicial-style opinion explaining its reasoning
What the AI Said (Excerpt)
One of the models concluded that federal sentencing should rely on a uniform national standard, not varying state laws:
“Option 1 best advances [the statutory] mandate because it anchors a federal sentencing enhancement to a single national drug schedule, rather than to fifty-plus sets of evolving state schedules.”
On another issue—whether courts should look at the law at the time of sentencing or at the time of the prior conviction—the model emphasized a core principle of fairness:
“The career offender guideline is, at bottom, a recidivist enhancement… That enhancement rationale is most coherent when the legal status of the prior conduct is evaluated as of the time it occurred.”
Why This Matters
What’s striking isn’t just that the AI reaches plausible legal conclusions—it’s that it does something Wroblewski argues the Commission often does not:
👉 It explains itself clearly, systematically, and transparently.
That raises a deeper question:
If AI can produce reasoned, consistent, and transparent sentencing policy analysis…
what exactly is the comparative advantage of human commissioners?
Read the Full Essay
This is just a small slice of a much longer and more detailed experiment—including full AI-written “opinions” on both circuit conflicts.
👉 You can read the full piece here:
Sentencing Matters-Jonathan Wroblewski
If this piece resonated with you, consider sharing it or leaving a comment. To support this work and help spread awareness about justice reform for white-collar defendants, subscribe to White-Collar Journal and stay connected. John DiMenna is a member of the White Collar Support Group.
To leave a comment, Substack may ask you to verify your email address (a one-time step to prevent spam). You don’t need to subscribe or create an account. Just check your inbox for a one-time link.